Julia package for selecting hardware, libraries, and algorithms for a computation
Author timholy
18 Stars
Updated Last
4 Months Ago
Started In
July 2016


Build Status

A Julia package for choosing resources (hardware, libraries, algorithms) for performing computations. It exports a set of types that allow you to dispatch to different methods depending on the selected resources. It also includes simple facilities that help package authors manage the loading of code in a way that depends on resource availability and user choice.


This package exports the following types of resources:

  • CPU1 (single-threaded computation)
  • CPUThreads (multi-threaded computation)
  • CPUProcesses (multi-process computation)
  • ArrayFireLibs (using the ArrayFire package)
  • CUDALibs (GPU computation using NVIDIA's CUDA libraries)
  • OpenCLLibs (GPU computation using the OpenCL libraries)

Algorithm selection is performed by passing a resource instance as an argument (conventionally, the first argument) to a function, where CPU1() is the typical default. Instances of these types may optionally store additional settings that you can customize; for example you could define a TimeOut type:

struct TimeOut

and then call some algorithm as

optimize(CPU1(TimeOut(3)), f, x)

As a package author, you could write optimize to check for the timeout value and terminate early once this time has been exceeded.

Usage as a user

Begin your session with

using ComputationalResources

Then choose any resources you have available, for example:


It's important to do this before you load any packages supporting ArrayFire-specific implementations; otherwise, calls such as

filter(ArrayFireLibs(), b, a, data)

are likely to throw a MethodError because the relevant specializations will not have been loaded.

If you'd like to make your selection of available resources automatic, you can add such lines to your .juliarc.jl file.

Usage as a package author

You can make the loading of code dependent upon the resources selected by the user. We'll use the "Dummy" package as an example (see also the test/packages folder for additional examples). This package could have the following file structure:


where ... means additional files. Dummy.jl might start like this:

module Dummy

using ComputationalResources

# You need an __init__ function that will manage the loading of
# sub-modules that implement specializations for different
# computational resources
function __init__()
    # Enable `using` to load additional modules in this folder
    push!(LOAD_PATH, dirname(@__FILE__))
    # Now check for any resources that your package supports
    if haveresource(ArrayFireLibs)
        # User has indicated support for the ArrayFire libraries, so load your relevant code
        @eval using DummyAF
    # Put additional resource checks here
    # Don't forget to clean up!

# Now define the methods you'll export, using single-threaded CPU
# computations as the "foundation"
function foo(resource::CPU1, args...)
    # awesome algorithm goes here

# Typically you should select a default resource
foo(args...) = foo(CPU1(), args...)


Your DummyAF module is implemented in DummyAF.jl, which might look like this:

module DummyAF

using ComputationalResources, Dummy, ArrayFire

function, args...)
    # specialization for the same computation, but using the ArrayFire libraries instead


Note that the ArrayFire package was loaded by DummyAF but not by Dummy; as a consequence, users who do not have this package installed will not experience any errors. Assuming it's optional, you should not add ArrayFire to your package's REQUIRE file.